A Framework for Accurate Time Series Classification Based on Partial Observation

摘要

Time series classification problems are solved using a variety of algorithms. The success of each technique is earmarked by measures of efficiency and accuracy. In order to achieve efficiency and accuracy, existing methods detect the relevant data in segments within a time series. Within this trend, we propose a framework for Detecting Optimal Partial Observation (DOPO) in time series classification. The framework developed is applicable to any time series database. It isolates the most relevant data for classification and eliminates excess data. This framework measures accuracy of partial observations by utilizing a 10 × 5 fold cross validation. In this paper, the DOPO framework is tested on all 85 data sets from the UCR Time Series Classification Archive. The time series classification method Dynamic Time Warping with 1-nearest neighbor is used with the framework. Our results indicate that the framework improved or maintained accuracy in a majority of tested data sets, while sensing only half the data.